Who Has NVIDIA's Blackwell GPUs: Market Size & Fragmentation (Q2 2026)

NVIDIA shipped approximately 3.2 million Blackwell GPU packages by end of 2025. Ten operators control 67 percent of Blackwell compute: five hyperscalers (Microsoft, Meta, Google, Amazon, Oracle) hold 49 percent, and five anchor neoclouds (CoreWeave, Nscale, Lambda, Nebius, Crusoe) hold 18 percent. The rest is spread across enterprises, sovereign AI, and smaller neoclouds.

NVIDIA has orders for another 4.3 million packages, which we estimate at 2.7 million after discounting for timing and cancellations.

3.2 million Blackwell packages

Blackwell is NVIDIA's current GPU architecture generation, succeeding the Hopper generation (H100, H200) that dominated 2023 and 2024. Where a single Hopper chip contains one silicon die, each Blackwell package fuses two dies into a single module, roughly doubling the transistor count and training throughput. When Jensen Huang said "six million Blackwell" shipped at his October 2025 GTC keynote, he was counting dies. The actual package count is approximately three million. ^{[1]Jensen Huang, GTC Washington DC keynote, session DC1204, October 28, 2025. Slide footnote clarified that the Blackwell architecture uses two GPU dies per chip; the "six million Blackwell" figure thus refers to dies, implying approximately 3 million complete packages.https://www.nvidia.com/en-us/on-demand/}

Blackwell ships in two main SKUs. The B200 is the earlier variant, shipping from mid-2025 in both the HGX server form factor (eight GPUs per board) and the NVL72 system (72 GPUs wired together in a rack). When a B200 is paired with NVIDIA's Grace CPU in the NVL72 rack, the combined module is sold as GB200; the standalone HGX version is just B200. The same convention applies to the B300/GB300. The B200 can be air-cooled but easily throttles without sufficient cooling.

The B300 is the higher-clocked successor and cannot be air-cooled. It began volume shipments around September 2025 and crossed over B200 in quarterly production by November 2025. ^{[2]NVIDIA CFO Colette Kress, Q3 FY2026 earnings call, November 19, 2025: "GB300 crossed over GB200" during that quarter; cumulative B200 still dominates through October.} The cumulative installed base through the end of 2025 skews approximately 75 percent B200 and 25 percent B300. ^{[3]TrendForce CY2025 Blackwell rack mix estimate of approximately 81 percent GB200 versus 19 percent GB300 (July 24, 2025).}

NVL72 racks account for approximately 2.0 million of the 3.2 million packages shipped. The remaining 1.2 million ship as HGX boards and DGX systems (NVIDIA's pre-integrated servers), which deploy in a wider range of configurations, from large training clusters down to single-server setups. ^{[4]Morgan Stanley supply chain research, NVL72 rack count analysis, 2025. At 72 packages per rack, yields 2.0 million packages in NVL72 form factor.}

From shipped to clustered

Roughly 10 percent of shipped packages sit in staging at any given time. The path from factory to operational cluster runs through installation, power commissioning, and network fabric integration, and typically takes 4 to 12 weeks.

Of the packages that are deployed, roughly a third run as standalone nodes or micro-clusters for inference, development, or edge workloads. The rest sit in clusters of 256 or more interconnected GPUs.

Who owns what

Clustered Blackwell compute is distributed across eight segments, attributed by owner-operator rather than end seller to prevent double-counting. When CoreWeave leases GPUs to Microsoft, that capacity counts in CoreWeave's segment.

Across all eight segments, roughly 125 operators hold clustered Blackwell compute: 5 hyperscalers, 5 anchor neoclouds, 22 to 40 enterprise deployments, 1 frontier AI lab (xAI), 20 to 40 Tier 2 cloud and marketplace providers, 8 to 12 small neoclouds, 6 to 8 sovereign programs, and 10 to 15 research, government, and academic sites. This operator count is only for operational clusters of 256 or more GPUs. When including smaller clusters, the number of total operators might be three times larger.

Cluster sizes

Twelve operational large clusters exist globally at December 31, 2025. A "large cluster" here is 10,000 or more Blackwell packages under a single operator, which may span multiple physical sites for fleet-aggregate operators like CoreWeave and Meta.

The long tail sits at the other end: 170 to 215 operational clusters below the $100 million hardware threshold (roughly 2,000 packages at $50,000 blended system-level ASP). Between the long tail and the large clusters sit roughly 140 medium clusters of 1,000 to 9,999 GPUs. The long tail is the hardest to size because enterprise, sovereign, and small-provider deployments are less observable than hyperscaler buildouts.

Projected vs. realized

Eighteen named pipeline deals total approximately 4.3 million announced Blackwell packages in commitments. Realization rates are the share of announced deals that actually deliver. Applying historical rates by buyer type to the pipeline produces an expected figure of approximately 2.7 million packages. ^{[5]NVIDIA reported approximately 3.6 million unit backlog across major cloud providers in December 2025 press coverage.}

Realization rates vary by buyer. Hyperscaler internal capex (GPUs funded from a company's own balance sheet) converts at 85 to 95 percent. Anchor neocloud offtakes (contracts where a single dominant customer commits to most of a cluster's capacity) convert at 70 to 85 percent. Sovereign AI programs convert at 40 to 70 percent.

Forward projections

Three inputs drive the 2026-2027 outlook: realization of the contracted pipeline, organic new orders, and expanding capacity at CoWoS-L. CoWoS-L is TSMC's advanced packaging process that fuses the two Blackwell dies onto a single substrate, and it is the primary supply bottleneck.

Cumulative shipments approximately double by end of 2026 and grow by another 2.4 million in 2027, while large clusters more than triple from 12 today to 40-60 by end of 2027. Rubin architecture begins shipping in H2 2026 per NVIDIA's disclosed production schedule, adding to Blackwell demand rather than replacing it.

Risks to realization

Five primary risks could compress the forward projections, ordered from most to least concerning.

Calculations

3.2M packages shipped

Three sources triangulate. Jensen Huang stated 3.0M packages cumulative through October 2025 at GTC Washington DC. ^{[1]Jensen Huang, GTC Washington DC keynote, session DC1204, October 28, 2025. Slide footnote clarified that the Blackwell architecture uses two GPU dies per chip; the "six million Blackwell" figure thus refers to dies, implying approximately 3 million complete packages.https://www.nvidia.com/en-us/on-demand/} Morgan Stanley channel checks counted 28,000 NVL72 racks shipped in CY2025, which at 72 packages per rack yields 2.0M in NVL72 form factor. ^{[4]Morgan Stanley supply chain research, NVL72 rack count analysis, 2025. At 72 packages per rack, yields 2.0 million packages in NVL72 form factor.} The remaining 1.2M ships as HGX and DGX boards. Epoch AI's chip sales model gives 3.4M median with a 90 percent confidence interval of 2.9 to 4.0M.

10 percent in staging

Q4 CY2025 was the peak shipment quarter at approximately 1.0M packages, of which 35 percent remained in staging or transit at December 31. Earlier quarterly cohorts deployed fully. The weighted lag across the installed base is approximately 10 percent.

1.9 million clustered (59 percent of shipped)

NVL72 systems are purpose-built for multi-rack clusters: 85 percent deploy in clusters of four or more racks. HGX and DGX systems are more varied: 35 percent deploy in 256-plus GPU clusters, with the majority in single-node inference, development, or edge configurations. Weighted: (2.0M × 0.85) + (1.2M × 0.35) = 2.12M in clusters before staging lag. After applying the 10 percent staging lag, roughly 1.9 million packages are operationally clustered, or 59 percent of the 3.2 million shipped.

49 percent hyperscaler share

Jensen Huang stated cloud service providers represent approximately 50 percent of data center revenue at the Q4 FY2026 earnings call. ^{[6]NVIDIA Q4 FY2026 earnings call, CFO commentary, February 25-26, 2026. Jensen Huang statement on data center revenue split by customer type: approximately 50 percent cloud service providers.} After accounting for 8 to 15 percent volume discounts on hyperscaler purchases, this maps to roughly 49 percent of clustered Blackwell units.

170 to 215 sub-$100M clusters

Approximately 48 percent are directly observed: Tier 2 cloud providers tracked by cloud-gpus.com (48 clusters), named small neoclouds (34), and anchor neocloud satellite sites (10). ^{[7]cloud-gpus.com, reserved 1-year B200 pricing as of April 2026. 31 providers tracked, 15 or more offering B200.https://cloud-gpus.com} Another 17 percent is partially observed through enterprise deployments at Block, IBM, Tesla, and smaller Mistral nodes. The remaining 35 percent is modeled from segment share, capturing enterprise residual, research and government clusters, sovereign small deployments, and regulated enterprise private cloud. ^{[8]SemiAnalysis, ClusterMAX 2.0, 209 GPU cloud providers tracked. Captures the small neocloud and Tier 2 cloud provider universe.} Omdia and Dell'Oro independent trackers confirm a defensible range of 178 to 215. Consulting firms frequently report 240 to 280 by conflating announced and contracted capacity with operational deployment.

Pipeline realization rates

Rates differ by buyer type:

• Hyperscaler internal capex: 85 to 95 percent. Board-approved, funded, with land and power secured. Meta Prometheus and Hyperion, Microsoft Fairwater, and AWS internal builds sit here.
• Anchor neocloud offtakes: 70 to 85 percent. Contractual obligations with named offtakers. CoreWeave backlog, Nscale-Microsoft, Lambda-Microsoft, and Nebius-Microsoft sit here.
• Sovereign AI: 40 to 70 percent. Subject to political cycles, permitting delays, and export license uncertainty. Historical midpoint of 55 percent based on prior procurement cycles.
• Frontier labs: 60 to 80 percent. Fundraising-dependent. xAI's Colossus 2 target sits here.
• Enterprise: 50 to 70 percent. ROI reassessment is common between announcement and deployment.

Saudi HUMAIN's 600,000-GPU announcement at 50 percent yields 300,000 expected. UAE G42 Stargate's 500,000 at 60 percent yields 300,000. Cross-checked against NVIDIA's reported 3.6M unit backlog (December 2025), the 2.7M realization-adjusted pipeline accounts for approximately 75 percent of backlog. ^{[5]NVIDIA reported approximately 3.6 million unit backlog across major cloud providers in December 2025 press coverage.} The remaining 0.9M sits in MOUs, LOIs, and earlier-stage commitments not captured as named pipeline deals.

Forward projection build

EOY 2026: 3.2M base plus approximately 2.2M organic net-new shipments (NVIDIA and Morgan Stanley guidance) plus approximately 1.3M from pipeline realization equals approximately 6.7M cumulative. Operational clusters approximately double from 275 to 400-500 as named pipeline sites energize.

EOY 2027: approximately 2.4M net-new shipments (including approximately 0.8M from remaining pipeline realization) brings cumulative to approximately 9.1M. The deceleration relative to 2026 reflects conservatism around Rubin transition timing. If Rubin launches on schedule and cannibalizes Blackwell Ultra orders, 2027 Blackwell-only figures would trend lower while total NVIDIA DC shipments remain on plan.

Supply constraints

TSMC CoWoS-L capacity expanded from approximately 75,000 wafers per month in 2025 to 120,000-130,000 in 2026 and 140,000-160,000 in 2027. NVIDIA secured approximately 70 percent of allocation. At Blackwell die-per-wafer yields of 25 to 30 and two dies per package, this supports approximately 13 to 16 million packages annually at the 2026 ceiling, well above the 2.2M net-new 2026 projection.

HBM3e supply from SK Hynix, Samsung, and Micron tightens at higher volumes. Grid power and permitting are the binding constraint for large-cluster deployment in 2026, with Virginia (Dominion Energy service territory) the tightest market due to moratorium pressure. Texas, Ohio, Wisconsin, and international deployments in the Nordics and Malaysia have capacity more readily available for 2026 energization.

References

1. Jensen Huang, GTC Washington DC keynote, session DC1204, October 28, 2025. Slide footnote clarified that the Blackwell architecture uses two GPU dies per chip; the "six million Blackwell" figure thus refers to dies, implying approximately 3 million complete packages.
2. NVIDIA CFO Colette Kress, Q3 FY2026 earnings call, November 19, 2025: "GB300 crossed over GB200" during that quarter; cumulative B200 still dominates through October.
3. TrendForce CY2025 Blackwell rack mix estimate of approximately 81 percent GB200 versus 19 percent GB300 (July 24, 2025).
4. Morgan Stanley supply chain research, NVL72 rack count analysis, 2025. At 72 packages per rack, yields 2.0 million packages in NVL72 form factor.
5. NVIDIA reported approximately 3.6 million unit backlog across major cloud providers in December 2025 press coverage.
6. NVIDIA Q4 FY2026 earnings call, CFO commentary, February 25-26, 2026. Jensen Huang statement on data center revenue split by customer type: approximately 50 percent cloud service providers.
7. cloud-gpus.com, reserved 1-year B200 pricing as of April 2026. 31 providers tracked, 15 or more offering B200.
8. SemiAnalysis, ClusterMAX 2.0, 209 GPU cloud providers tracked. Captures the small neocloud and Tier 2 cloud provider universe.